The room and the question
Every second Thursday, Concord Software Developers meets at Flywheel Coworking in downtown Concord. The July session ran almost two hours and never really left one subject: AI agents — the kind that don't just suggest code but go do things. What follows are the ideas from that room, attributed to the group rather than to individuals, because that's the deal when the recorder is on.
The short version: these are working engineers, not evangelists or skeptics, and the conversation was more practical than anything you'll read on a keynote stage.
Agents are already doing real work
The session opened with a live demo: a single high-level prompt, and an agent found the hosting platform's CLI documentation through web search, configured the environment, and deployed an application on its own. No step-by-step babysitting.
The detail that stuck with us was why it worked: the platform deliberately optimizes its documentation so AI agents can find it. Think about what that means. Search optimization used to be about people finding your website; now whether an AI tool can use your product depends on whether an AI can find your docs. That's the answer-engine shift we keep telling clients about, showing up in a place nobody was watching.
The security conversation got serious
A big stretch of the evening went to agent "skills" — shareable add-ons you can install to extend what an agent can do. The group's consensus was blunt: installing an unvetted skill from a marketplace is the same category of decision as running an untrusted command on your machine. It can read what your agent can read.
The mitigations people actually use were the useful part: dedicated machines with no personal data on them, segmented networks so a misbehaving agent can't reach everything, separate scoped accounts and API keys for agents instead of your own credentials, and having one AI review a skill's code before another AI is allowed to use it. And one rule came up repeatedly: don't let agents install their own tools. A human vets what goes in.
The advice that stuck with us
During a deep-dive on database choices, someone offered the most transferable line of the night: pick the solution that buys you time. If the simple option can carry you for six months, take it — you'll learn whether the product deserves the complex option before you've paid for it. That applies to a lot more than databases.
The other thread worth stealing was about getting automation adopted at work: translate the time saved into dollars before you pitch it, put the proposal in writing because a document commands attention a chat message doesn't, and pilot small to earn trust before asking for the full rollout. If the idea gets ignored, bring it up again. Persistence was described as a feature, not a flaw.
You're still the architect
The night ended somewhere philosophical: does leaning on AI make developers worse? The room's answer was no — but it makes unaccountable developers worse, faster. When AI output fails, the cause is usually a human who didn't provide enough context, the same way a junior developer fails when they're managed badly. One phrase that landed: the danger isn't the tool being dumb, it's you dropping your guard because the tool sounds smart.
The developer is still the architect, still responsible for what ships, and rigorous review of AI-written work isn't optional. That matches what we see in our own client work every week.
Come sit in
Concord Software Developers meets every second Thursday at 6:30pm at Flywheel Coworking, 57 Union Street South, downtown Concord. It's free, it's honest, and the hallway conversations are as good as the sessions. Bring what you're building.